AI News

One-Step Generative Modeling via Training Dynamics Action (arxiv.org)

Beyond Affine Transformations: A Soft Dominance Layer for Coordinate-Wise Neural Computation (arxiv.org)

Evaluating Hybrid Quantum-Classical Models for Reduced-Order Brain Deformation Dynamics (arxiv.org)

Denoising Surface: Modeling and Predicting Inference Cost for Diffusion LLM Serving (arxiv.org)

Gumbel Straight Flow: Distilling Autoregressive Models into One-step Flow Maps (arxiv.org)

Score the Update, Not the Token: Descent-Aligned Routing for Combinatorial LoRA Experts (arxiv.org)

Redundancy Meets Synergy: Dependency-aware Expert Selection for MoE via Submodular Optimization (arxiv.org)

Exact information accounting for SGD methods (arxiv.org)

One pool, many targets: a conservation layer and what archival data can identify (arxiv.org)

Every Batch Is Its Own Validation Set: Leave-One-Out Gradient Matching for Online Data Selection in LLM Fine-Tuning (arxiv.org)

XOR-Trellis: Ultra-Low-Complexity Dequantization and Curvature-Aware Hadamard-Free LLM Quantization (arxiv.org)

Smoother Flow Matching via Contrastive Trajectory Repulsion (arxiv.org)

The Life Cycle of a Massive Activation: Stochastic Birth, Weight-Decay-Driven Growth, and Competitive Consolidation (arxiv.org)

CommunityKV: Efficient Long-Context Decoding via Graph Partitioning (arxiv.org)

Source Identification Is Not Fitness Testing: Measuring the Limits of Synthetic-Data Attribution (arxiv.org)

Flowing Faster to Coordinate: One-Step Online Multi-Agent Flow Policies (arxiv.org)

Spatial Lifting for Dense Prediction (arxiv.org)

Emergent Object Binding Has a Finite Spatial Horizon (arxiv.org)

STATERA: Hidden Mass Estimation via Zero-Shot Sim-to-Real Kinematics using Frozen Temporal Tubelets (arxiv.org)

New Snake-in-the-Box Records via Snakepit Surgery and Learned Construction (arxiv.org)

TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-Tuning (arxiv.org)

IrekoGPT: Turning Structured Pruning into Post-Hoc Slimmable LLMs (arxiv.org)

FERPO: Forward Entropy-Regularized Policy Optimization (arxiv.org)

Cost-augmented Schr\"odinger bridges on graphs are exactly solvable: a Feynman-Kac tilt replaces learned control (arxiv.org)

The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models (arxiv.org)

Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning (arxiv.org)

Optimal Transport Meets Reinforcement Learning: A Survey (arxiv.org)

Measuring the Microtask Eligibility Gap: When Is an Off-the-Shelf SLM Enough for an Agent Harness? (arxiv.org)

Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features (arxiv.org)

Higher-Order Molecular Grammars for Generative and Foundation Models in Chemistry (arxiv.org)

Decoding Looped Transformers Better for (Almost) Free (arxiv.org)

Agent Evaluation Reliability: More Tasks Won't (Always) Fix An Agent Leaderboard (arxiv.org)

SoftServe: A Scalable Quasi-Newton Method for Deep Learning (arxiv.org)

From Gradients to Capabilities: Understanding Multi-Teacher On-Policy Distillation (arxiv.org)

Effective Resistance and Graph Neural Network Reliability in Tissue-Specific Interactomes (arxiv.org)

Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair (arxiv.org)

Faynt: Scaling and Optimizing Policies for Competitive Melee (arxiv.org)

Finetuning with Sampling: SFT Learns Better Than You Think (arxiv.org)

Linear Programming Representations and Strongly Polynomial Algorithms for Robust Markov Decision Processes (arxiv.org)

Local Support Learning (arxiv.org)

Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability (arxiv.org)

Learn the Directions, Normalize the Gains: Post-Training Normalization for LoRA (arxiv.org)

Encoded but Disconnected: Decomposing Vision-Language Model Failures under a Patching Null (arxiv.org)

Foundations without Fundamentals: Zero-Shot Blind Spots in Time Series FMs (arxiv.org)

Distributionally Robust Schr\"odinger Bridge (arxiv.org)

On Evaluating Quantum Kernel Robustness for Low-Resource Cross-Corpus Audio Deepfake Detection (arxiv.org)

CARM: Cancellation-Aware Response Masking for LLM Reinforcement Learning (arxiv.org)

Relative Transitions, Not Absolute Destinations: A Transfer-and-Ground Framework for Target-Trajectory-Free Human Mobility Generation (arxiv.org)

On Language Drift during RLVR Post-Training (arxiv.org)

BranchIP: Learning Adaptive Equivariant Computation for Interatomic Potentials (arxiv.org)